Compare · As of 2026-10-03

How ALPHA-Memory Compares

ALPHA-Memory is in development and not in use yet. This page sets its design next to memory products a team can use with Claude Code today, row by row. It compares features and facts; it is not a benchmark and not a ranking. What each status word means, and where the cells come from, is under How We Compare below the table.

The Comparison

How We Compare

Every cell starts with one status word:

Shipped

Generally available today.

Beta

Available under the vendor's own beta or preview label.

Measured

Our column only: measured offline in our own test, not a running service.

Planned

For us: in the design, not built or not in use. For another product: on its public roadmap.

No

The vendor says so, or a search of its public code found nothing while a control string in the same code was found.

Unknown

No public source settles it, or the cell is disputed and being re-checked. Never turned into No.

N/A

The row does not apply to this product; the cell says why.

• Every row asks about the product as it is used with Claude Code. A cell of the quality row that rests on an independent benchmark says how the benchmark ran the product.
• Every cell of another product links its source and shows the date we last read it. The date at the top of the page is the oldest of those dates. A source is the vendor's own website or documentation, or its public repository at a fixed commit; a cell about an independent benchmark cites the benchmark's paper.
• Quotes are verbatim, except that long dashes are shown as hyphens. A quote from a table or a pricing card gives its cells in order; a quote from code is one or more lines of it, joined, without their comment markers.
• The cells of the other products rest on public sources: documentation, public code and, where a cell cites an independent benchmark, that benchmark's paper. We did not sign up for or test any of the hosted products. Claude Code is the tool we work in; its cells, too, cite only public sources.
• No benchmark figures yet. A number that compares us with a named product appears here only once we have measured both sides the same way; we do not repeat a vendor's own benchmark results.
• Our own column rests on internal documents and measurements that you cannot yet inspect; each cell says so, or links the section of this site it rests on.
• Nothing of ours is in use yet: wherever another product reads Shipped and our cell reads Planned or No, that product is ahead of us today. The rows where our design itself has nothing to show yet stay in the table.

Each Row, With Its Sources

What each row asks and, for every product, the status, the detail, the source, with its quote where the source states the fact, and the date we read it.

Status

Status, with an as-of date.

• ALPHA-Memory · In Development Planned In development; not in use yet. This site: Where It Stands. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Built into Claude Code; on by default locally. Auto memory is on by default in local sessions. Vendor documentation, checked 2026-10-04.
• Supermemory Shipped Coding plugins available on every plan. Coding plugins are available on every plan, including Free. Vendor documentation, checked 2026-10-03.
• Hivemind Shipped Self-serve install and sign-in. The installer detects every supported assistant on your machine (table below), wires up the hooks, and shows a one-line consent prompt before opening a browser for sign-in. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Self-serve signup; free tier. Sign up and test the Platform with our free tier. No credit card required. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped Docs call MCP connection the "production product". production product is the MCP connection used from the assistant you already have. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped CMEM Cloud live; Team seats by reservation. CMEM Cloud is live. Get your private MCP link, sync your first observations, and never let your agent forget again. Vendor documentation, checked 2026-10-04. The public repository labels the server runtime behind the team features beta, and the home page offers the Team plan through "Reserve team seats". The page of the vendor's TeamBrain pilot (cmem.ai/teambrain) says: "Shared team memory, team access controls, and team onboarding are the pilot direction." and "Individual Pro accounts do not automatically share memories with teammates."
• Zep Cloud Shipped Self-serve plans available. Self-serve plans are credit based. Enterprise plans add governance, deployment options, and negotiated rates. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped Available today; starts with a 7-day trial. Basic Memory Teams is available today. Bring your existing notes and projects with you. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped Self-serve signup. Click Sign Up and create your account Vendor documentation, checked 2026-10-03.

Shared Across Developers

Developer B's agent recalls what developer A's agent learned, each developer with their own identity.

• ALPHA-Memory · In Development Planned Shared per project; one identity per developer. Internal design document (SPEC F6, D37), not public. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Version control: CLAUDE.md, subagent notes; auto memory machine-local. These instructions are shared with your team through version control, so focus on project-level standards rather than personal preferences. Vendor documentation, checked 2026-10-04. The quote is about CLAUDE.md. The memory page calls auto memory machine-local. One part of it can be shared: a subagent given the memory field at project scope keeps its notes in the project's .claude/agent-memory/ directory, which the vendor calls "shareable via version control" (code.claude.com/docs/en/sub-agents#enable-persistent-memory).
• Supermemory Shipped Organization member seats; Pro plan and higher. Your organization has a plan that supports team invitations. Pro and higher include this feature. Vendor documentation, checked 2026-10-03.
• Hivemind Shipped Shared workspaces; teammates invited per workspace. Invite teammates to the workspaces they need. Revoke in one click. Audit everything. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Shared project pool per repository; user_id per developer. Everyone on the repo reads and writes the same pool. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped Shared workspace; members invited by email, with roles. A decision your colleague recorded on Tuesday can answer your question on Thursday-in whichever connected assistant you use. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Beta Server (Beta): shared team pool, per-developer API keys. Each developer gets their own api key (so revocation is per-person). Vendor documentation, checked 2026-10-03.
• Zep Cloud Shipped Per-user sign-in; shared Context Graphs, project-wide. Connections initially use project-wide authorization: every authenticated MCP user can read every shared Context Graph in the token-bound project. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped Team workspace; per-seat members with roles. Anything a teammate writes is immediately available to everyone else and to their agents: a decision captured in one person's Claude session becomes context for another's Codex run. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped Organization members with roles share memory banks. Hindsight Cloud supports team collaboration with role-based access control. Vendor documentation, checked 2026-10-03.

How Memory Reaches Claude Code

How memory reaches Claude Code: every prompt, session start, or the agent pulls it.

• ALPHA-Memory · In Development Planned Every prompt, through a hook; tools on demand. Internal design document (SPEC F4, F5), not public. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Index at session start; topic files on demand. The first 200 lines of MEMORY.md, or the first 25KB, whichever comes first, are loaded at the start of every conversation. Vendor documentation, checked 2026-10-04. The same page says that topic files are not loaded at startup and that "Claude reads them on demand using its standard file tools when it needs the information." It also says that this limit "applies only to MEMORY.md" and that "Claude Code loads a CLAUDE.md file of up to 4 MiB in full". Claude Code's changelog also names a memory recall step; see the row Nothing Injected When Nothing Is Relevant.
• Supermemory Shipped Session start, every prompt, and MCP tool. Supermemory hooks: SessionStart loads project memories; UserPromptSubmit recalls memories relevant to the prompt; PreToolUse approves read-only Supermemory tool calls; Stop saves new conversation content to memory. Public code, checked 2026-10-03.
• Hivemind Shipped Session-start instructions, then agent-run grep on memory. DEEPLAKE MEMORY: You have TWO memory sources. ALWAYS check BOTH when the user asks you to recall, remember, or look up ANY information: Public code, checked 2026-10-03.
• Mem0 Platform Shipped First prompt, up to five memories; then pull. If the first prompt has at least 20 characters, the plugin searches with that prompt and injects up to five relevant memories. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped Hook injects context on every user message. Injects ContextStream context (preferences, lessons, rules) on every user message. Public code, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped Session-start push (default 50 observations); MCP/skill search. Queries SQLite for recent observations (configurable, default 50) Vendor documentation, checked 2026-10-03.
• Zep Cloud Shipped Pulled by the agent via MCP tools. The Context MCP Server lets an end user connect an MCP client (Claude, ChatGPT, Claude Code, Codex, Cursor, and others) to their own agent memory in Zep. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped Session-start brief via plugin; MCP tools. When a session begins, the plugin queries Basic Memory for your active tasks and recent work and puts a short brief in front of Claude Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped Session-start page roster; first-prompt synthesis; MCP tools. AUTO-INJECT once per session, on the first prompt Public code, checked 2026-10-04. Describes the Coding Agents plugin. The older Claude Code plugin in the same repository, which recalls before every prompt, is marked deprecated by the vendor.

Subagent Sessions Captured

Subagent work captured. Shipped: the product records a subagent's own transcript or tool calls, or its result as returned to the main agent, or the vendor documents a memory of its own for subagents; a setting the vendor documents for subagents counts, and the cell then says Opt-in. No: the product records sessions and leaves subagent work out. N/A: it records no session, and the agent writes each memory through a tool. Unknown: the public sources do not settle it.

• ALPHA-Memory · In Development Planned Main and subagent sessions. Internal design document (SPEC F1, D33), not public. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Opt-in: each subagent keeps its own notes directory. The memory field gives the subagent a persistent directory that survives across conversations. Vendor documentation, checked 2026-10-04. Auto memory is notes Claude writes, not a recording of a session. With the memory field set, a subagent writes its own notes into a separate directory.
• Supermemory Unknown Tool results off by default; background results unsettled. includeTools: [], Public code, checked 2026-10-04. The Stop hook uploads user and assistant text from the main transcript. A tool result, which is how a foreground subagent's answer returns to the main agent, is included only for tools listed in includeTools, empty by default. Claude Code's documentation (code.claude.com/docs/en/sub-agents) says that in an interactive session subagents run in the background by default, and that a background subagent's result reaches the main agent as a completion notification. The plugin's text filter (plugin/hooks/lib/transcript.js, lines 183 to 190, at the same commit) removes only system reminders and the plugin's own context block, so a notification written into the transcript as a user message would be uploaded with the user text. We did not run the product, so this is not settled. A search of the plugin at this commit for SubagentStop, subagent, sidechain and notification found nothing; the control string includeTools was found.
• Hivemind Shipped Subagent tool calls and responses captured. Subagent tool calls and responses Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Subagent results, as included in the main transcript. Subagent results included in the main transcript can provide supporting evidence for extraction; the main agent's final response establishes the outcome. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped SubagentStop hook installed for Plan subagents. event: ClaudeHookEvent::SubagentStop, matcher: Some("Plan"), hook_name: "subagent-stop", timeout: 15, Public code, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped Subagent tool calls captured via PostToolUse; summaries skipped. Every tool call by every subagent fires a `PostToolUse` hook, so claude-mem creates one provider-analyzed observation per call Vendor documentation, checked 2026-10-03.
• Zep Cloud N/A No capture; the agent writes through MCP tools. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams No Claude Code plugin registers SessionStart and PreCompact only. Basic Memory bridge hooks - brief Claude from the knowledge graph on session start, checkpoint the session before compaction. Public code, checked 2026-10-04. A search of the Claude Code plugin (plugins/claude-code) and of the hook command (src/basic_memory/cli/commands/hook.py) at this commit for SubagentStop and sidechain found nothing; the control string PreCompact was found. In those files the word subagent appears only in the plugin's design note (plugins/claude-code/DESIGN.md), on three lines: one describes Claude Code's own per-subagent memory, two list subagent memory as an idea to explore. The checkpoint before compaction keeps only user messages, shortened: the opening request and the last three. The transcript reader skips only meta lines and tool results (src/basic_memory/cli/commands/hook.py, line 940, at the same commit), and the checkpoint cuts each of the last three user messages to 200 characters (line 1071). Claude Code's documentation (code.claude.com/docs/en/sub-agents) says a background subagent's result reaches the main agent as a completion notification; if Claude Code writes it into the transcript as a user message, it is not filtered, so at most its first 200 characters could be one of those three. The checkpoint records no subagent transcript, no tool call and no full result. We did not run the product.
• Hindsight Cloud Shipped Opt-in via superseded plugin; current plugin drops subagents. The /hindsight-memory:create-agent skill creates a Claude Code subagent backed by its own Hindsight bank. Vendor documentation, checked 2026-10-04. The quote is about the older Claude Code plugin (hindsight-integrations/claude-code), which the same page says "is superseded by the Coding Agents plugin" and of which it adds "This page and the published package still work; they are no longer developed." That plugin can also keep tool results (its retainToolCalls setting, off in the settings.json it ships). The Coding Agents plugin that replaces it has no such skill and leaves subagent work out: it drops sidechain lines (hindsight-integrations/coding-agents/src/core/transcript.ts, lines 78 to 99, of the public repository vectorize-io/hindsight at commit f7dd3f4fd7420f7beec60c32c965e5e5cf7be066: "Drops thinking blocks, isMeta/isSidechain lines, injected memory, and empty turns.") and, for Claude Code, every tool result, which is how a foreground subagent's answer returns to the main agent. It also removes task-notification blocks from user text (hindsight-integrations/coding-agents/src/core/transcript-util.ts, lines 13 to 21, at the same commit), which that file describes as how "Claude Code reports a background task's outcome"; Claude Code's documentation (code.claude.com/docs/en/sub-agents) says a background subagent's result reaches the main agent as a completion notification. A search of hindsight-integrations/coding-agents at that commit for SubagentStop and create-agent found nothing; the control string isSidechain was found.

Redaction Before Upload

Redaction on the PC before upload, and which layers it uses.

• ALPHA-Memory · In Development Planned Five layers; exact known values first. This site: What Leaves Your PC. Checked 2026-10-03.
• Claude Code Built-In Memory N/A Auto memory: local files, no memory-store upload. Vendor documentation, checked 2026-10-04. Claude Code Projects (public beta on Pro and Max plans) have a project memory of their own, which the vendor calls "separate from the auto memory Claude Code keeps on your machine" (code.claude.com/docs/en/claude-projects).
• Supermemory No Claude Code plugin has no redaction step. .replace(/<system-reminder>[\s\S]*?<\/system-reminder>/g, '') Public code, checked 2026-10-04. The quoted line is from the function that cleans text before upload: it removes system reminders and the plugin's own context block. A search of the plugin at this commit for redact, mask and scrub found nothing (sanitize appears only for repository names and status events); the control string MIN_SIMILARITY was found.
• Hivemind Shipped Token patterns, secret key names, entropy backstop. ~30 known provider token schemes (keep the scheme prefix as a hint). Public code, checked 2026-10-03.
• Mem0 Platform Shipped Credential patterns, redacted before anything is sent. Values that look like credentials are redacted before anything is sent. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped Environment files, keys, certificates excluded before upload. Ignore rules, environment files, keys, and certificates are excluded before anything is indexed. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped Opt-in regex patterns before storage; <private> tags. Opt-in regex-based detection that replaces common secret patterns with <redacted/> placeholders before storage Vendor documentation, checked 2026-10-03.
• Zep Cloud N/A No local component; remote MCP server only. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams No Hook redaction subsystem removed in v0.23.0. Removed the hook-specific redaction subsystem Vendor documentation, checked 2026-10-03.
• Hindsight Cloud No Not on the PC; server redacts before storage. Memory Defense screens every retain call before content reaches storage, so secrets, prompt injections, and tampering attempts never enter your memory bank. It is always-on for Hindsight Cloud Vendor documentation, checked 2026-10-04. The row asks for redaction on the PC. A search of the public repository vectorize-io/hindsight (folder hindsight-integrations/coding-agents) at commit f7dd3f4fd7420f7beec60c32c965e5e5cf7be066 for redact, mask and scrub found no step that masks secrets (redact appears in three comments about a rewritten transcript and as the package name @npmcli/redact in package-lock.json; mask appears twice, about configuration precedence); the control string isSidechain was found. Of the server side the quoted page adds: "Each detector only runs if its rule is present in the bank policy and its entitlement flag is enabled for the org."

Who Can See a Memory

Who can see a memory, and whether recall follows each developer's access to secrets and projects.

• ALPHA-Memory · In Development Planned Project by default; private by explicit act. Internal decision record (D37), not public. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Auto memory machine-local; project CLAUDE.md via source control. Auto memory is machine-local. All worktrees and subdirectories within the same git repository share one auto memory directory. Files are not shared across machines or cloud environments. Vendor documentation, checked 2026-10-04. See the row Shared Across Developers for subagent notes kept in the repository.
• Supermemory Shipped Members see all containers; per-tag limits on Scale. For a Member, keep Full access to include all containers. On Scale or Enterprise, choose Restricted to limit access to selected container tags. Vendor documentation, checked 2026-10-03.
• Hivemind Shipped All workspace users can read captured data. All users in your Deeplake workspace can read this data. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Project memory: everyone on repo; personal: private. Everyone on the repo reads and writes the same pool. Project memory never carries a `user_id`, so teammates' searches never mix in your preferences. Vendor documentation, checked 2026-10-03.
• ContextStream Shipped Workspace roles; personal records visible only to owner. Personal records are visible only to their owner, and on Team plans conversation transcripts are personal by default. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Beta Reads bounded by API key's team, optional project. An API key carries a **team** (always) and an optional **project** scope; that scoping bounds every read, write, and delete. Vendor documentation, checked 2026-10-03.
• Zep Cloud Shipped Own user graph; shared graphs project-wide or ABAC. Connections can use project-wide access or UserGroup ABAC to grant specific graphs to each user. Plans without ABAC retain project-wide access. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped Per project: Team or Invite-only; role-based. Every project in a team workspace has an access level, and every member's reach into it follows their role. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped Role-based access control per organization. Hindsight Cloud supports team collaboration with role-based access control. Vendor documentation, checked 2026-10-04. By default an API key has access to every memory bank of the organization, and it can be limited to chosen banks (docs.hindsight.vectorize.io/api-keys): "You can restrict a key to only access specific banks".

Store Region

Store region stated.

• ALPHA-Memory · In Development Planned EU store; model calls leave the EU. This site: What Leaves Your PC. Checked 2026-10-03.
• Claude Code Built-In Memory N/A Auto memory: local files, no hosted store. Vendor documentation, checked 2026-10-04. Claude Code Projects (public beta on Pro and Max plans) have a project memory of their own, which the vendor calls "separate from the auto memory Claude Code keeps on your machine" (code.claude.com/docs/en/claude-projects).
• Supermemory Unknown No region stated; privacy policy names Timescale. Timescale for database services Vendor documentation, checked 2026-10-03.
• Hivemind Unknown No region stated for Hivemind Cloud; BYOC offered. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Amazon Web Services, US (subprocessor list). Amazon Web Services • Cloud provider US Infrastructure Hosting Vendor documentation, checked 2026-10-04. The same list names TurboPuffer as "Vector DB" and gives no location for it.
• ContextStream Shipped OVHcloud; EU-region storage on Team and Enterprise. Standard DPA. EU-region storage available on Team and Enterprise. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Unknown Sync hub: Fly region iad; database region unstated. primary_region = "iad" Public code, checked 2026-10-04. The same file names the database provider (Neon) in its first line, without a region.
• Zep Cloud Shipped us-west-2; EU residency on request, Emerging program. Customer data remains within us-west-2, Zep Cloud’s region. Vendor documentation, checked 2026-10-04. Zep's Emerging Companies program page (https://www.getzep.com/emerging, read 2026-10-04) states: "The program runs on Zep Cloud in the US by default, with EU data residency available on request."
• Basic Memory Teams Shipped Servers located in the United States. Basic Memory Cloud servers are located in the United States. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Unknown No region stated. Vendor documentation, checked 2026-10-03.

Extraction Model

Which model turns sessions into memories. Shipped: the vendor names it, or documents that the customer chooses it. Unknown: the public sources do not settle it. N/A: no model is called.

• ALPHA-Memory · In Development Planned Claude Sonnet 5.5, chosen by us. Internal decision record (D33), not public. Checked 2026-10-04.
• Claude Code Built-In Memory Unknown Claude saves the notes; model not named. Vendor documentation, checked 2026-10-04. The memory page says Claude writes the notes as it works in the session, and names no model. Which model a session runs is the customer's choice (the /model command, code.claude.com/docs/en/model-config). Claude Code's changelog (CHANGELOG.md, line 1342, of anthropics/claude-code at commit 1c229fcd1e1e4e452e29a8f116b45fe4cfe2c528) also names a "memory extraction" step (version 2.1.273) and does not say which model runs it, so the sources do not settle which model writes a memory.
• Supermemory Shipped Vendor's own model, named learner-1. Our model learner-1 helps any model continuously improve by extracting and dreaming on the context of any user, task, or tenant Vendor documentation, checked 2026-10-03.
• Hivemind Shipped Haiku via each developer's own claude -p. const CLAUDE_BASE_FLAGS = ["--no-session-persistence", "--model", "haiku"] as const; Public code, checked 2026-10-03.
• Mem0 Platform Unknown Not named; extraction uses platform default. Vendor documentation, checked 2026-10-03.
• ContextStream Unknown Not named; own key documented for generated answers. If you bring your own model key, those requests follow your provider’s terms. Vendor documentation, checked 2026-10-04. The security page says this of requests made when a feature or managed agent generates an answer; it does not say which model extracts memories. The vendor's page on own keys for its optional coding agent (contextstream.io/docs/contextcode/your-own-keys) says: "Memory, grounding, search, and every other feature work exactly the same either way."
• claude-mem / CMEM Cloud Shipped Customer picks: claude-mem observer, OpenRouter, Gemini, Anthropic plan. After sign-in you pick your memory provider - the claude-mem observer, your own OpenRouter or Gemini key, or your Anthropic plan. Vendor documentation, checked 2026-10-03.
• Zep Cloud Shipped Customer-chosen via BYOM, Enterprise add-on. Bring Your Own LLM (BYOM) lets you use your own accounts with model providers such as OpenAI, Anthropic, and Google. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams N/A Extractive checkpoint in hook; no LLM call. Public code, checked 2026-10-03.
• Hindsight Cloud Shipped Hindsight-hosted LLM, unnamed; BYO-LLM banks. the LLM is the one running on Hindsight-hosted infrastructure for managed banks, or your configured provider for BYO-LLM banks Vendor documentation, checked 2026-10-04. The Hindsight Cloud docs name BYO-LLM banks in this one sentence of the Memory Defense FAQ; they do not say on which plan. The self-hosted open-source server lets the operator choose the LLM provider (README.md, line 82, of vectorize-io/hindsight at commit f7dd3f4fd7420f7beec60c32c965e5e5cf7be066).

Nothing Injected When Nothing Is Relevant

Injects nothing when nothing is relevant. N/A: nothing is injected automatically at prompt time by default (a fixed load at session start, or the agent pulls). Shipped: an automatic prompt-time injection that code or documentation shows is skipped when nothing matches. Unknown: an automatic injection whose skip cannot be read.

• ALPHA-Memory · In Development Planned Tested configuration has no relevance floor. Internal design document (SPEC F4) and measurement, not public. Checked 2026-10-03. Nothing of ours runs in anyone's sessions yet. In our offline test on recorded sessions, the chosen configuration added memory to all 579 prompts, because it has no relevance floor yet. A floor is planned.
• Claude Code Built-In Memory Unknown Recall step in changelog; no-match behaviour not stated. Fixed session titles, memory recall and prompt hooks failing on Mantle or behind gateways that reject structured outputs Vendor documentation, checked 2026-10-04. The quoted changelog entry (version 2.1.288) names a memory recall step; version 2.1.284 speaks of "recalled memory notes before they reach Claude". The memory page (code.claude.com/docs/en/memory) says topic files are not loaded at startup and are read on demand, and describes a "Recalled 2 memories" message as Claude reading from its memory directory. None of them says what the step adds when nothing is relevant.
• Supermemory Shipped 0.55 similarity floor; one-time notice on first miss. const MIN_SIMILARITY = 0.55; Public code, checked 2026-10-03.
• Hivemind N/A Recall is pulled by the agent. Public code, checked 2026-10-03.
• Mem0 Platform Shipped Injects nothing when search returns no memories. if not result.memories: return {} Public code, checked 2026-10-04. The plugin's first-prompt search sends no threshold of its own; Mem0's search API documents a default similarity threshold of 0.1 (docs/api-reference/memory/search-memories.mdx, line 30, at the same commit).
• ContextStream Unknown Hook always writes a block; server supplies context. write_editor_output(editor, Some(combined))?; Public code, checked 2026-10-04. The hook's request to the service carries the workspace, project and session ids, the hook type, the folder path and the editor, and no prompt text (the same file, lines 455 to 475), so the public code does not show a relevance decision.
• claude-mem / CMEM Cloud N/A Session-start push; prompt-time injection opt-in, off by default. CLAUDE_MEM_SEMANTIC_INJECT: 'false', Inject relevant past observations on every UserPromptSubmit (experimental, disabled by default) Public code, checked 2026-10-04. By default memory arrives at session start: the most recent observations (default 50). With the experimental setting switched on, the prompt hook adds context only when its search returns observations (src/cli/handlers/session-init.ts, lines 287 to 309, at the same commit).
• Zep Cloud N/A Pull only; the agent decides when to search. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams N/A Status-based session brief; agent pulls by search. Public code, checked 2026-10-03.
• Hindsight Cloud Unknown Empty answer skipped; no-match reply requested as text. if (reflectAnswer && reflectRanThisTurn) blocks.push(buildSystemInjection(reflectAnswer)); Public code, checked 2026-10-04. The hook injects the first-prompt synthesis unless it comes back empty; the same file says of an empty answer, in lines 219 to 220: "a source can legitimately have nothing to say on a sparse bank". The query the hook sends asks for text when nothing matches (hindsight-integrations/coding-agents/src/core/inject.ts, line 31, at the same commit): "If the bank holds nothing that bears on the goal, say so in one line." What the service returns in that case was not tested.

When the Service Is Down

Behaviour when the service is down, with the hook timeout.

• ALPHA-Memory · In Development Planned Fails open after 3 s; miss logged. This site: FAQ. Checked 2026-10-03.
• Claude Code Built-In Memory N/A Auto memory: local files, no service or hook. Vendor documentation, checked 2026-10-04. Claude Code's changelog also names a memory recall step; see the row Nothing Injected When Nothing Is Relevant. Claude Code Projects (public beta on Pro and Max plans) have a project memory of their own, which the vendor calls "separate from the auto memory Claude Code keeps on your machine" (code.claude.com/docs/en/claude-projects).
• Supermemory Shipped Fails open; recall request capped at 4 s. Requests default to 3s; prompt recall allows 4s. Callers treat failure as "no memory this time", not a blocker. Public code, checked 2026-10-04. The prompt hook itself has a 5 s timeout: plugin/hooks/hooks.json, line 21, at the same commit.
• Hivemind Shipped Fails open; every hook exits 0 on error. Every hook exits 0 on error. A crash or timeout in any hook must never break the user's session. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Fails open; 6 s prompt-hook timeout. Mem0 never blocks normal Claude Code work when a hook fails. Vendor documentation, checked 2026-10-04. The timeout is in integrations/claude-code-plugin/hooks/hooks.json, line 22, at the same commit. Inside that hook the memory search has a 2 s timeout (core/hook_runner.py, line 75) and is sent a second time after a timeout or a connection failure (core/memory_core.py, lines 1845 to 1855).
• ContextStream Shipped Static reminder fallback; 5 s prompt-hook timeout. Falls back to a static reminder if the API is unreachable. Public code, checked 2026-10-04. The timeout is in crates/mcp-server/src/setup/hooks.rs, lines 1139 to 1144, at the same commit. The hook's request for context times out after 2 s (the cited file, line 483).
• claude-mem / CMEM Cloud Shipped SessionStart context hook exits 0; timeout 60 s. exit 0; }", "timeout": 60 Public code, checked 2026-10-03.
• Zep Cloud N/A No hook; remote MCP server only. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped Fails open; 20 s SessionStart hook timeout. Active harness verbs (session-start, pre-compact) are fail-open: any error logs to stderr and exits 0 - a hook must never disrupt an agent session. Public code, checked 2026-10-04. The 20 s timeout is in plugins/claude-code/hooks/hooks.json, line 10, at the same commit. In the cited file each query of the session brief has a 10 s budget (lines 78 to 79), and when every query of the primary project fails, the brief of a configured project says that it could not be read (lines 754 to 766).
• Hindsight Cloud Shipped Fails open; 30 s prompt-hook timeout. Failures never break the agent Vendor documentation, checked 2026-10-04. The 30 s timeout is stated on line 602 of the same file.

Quality Measured on Coding Sessions

Retrieval quality measured on coding sessions, with the method public: by the vendor, or by an independent benchmark that ran the product named in the column.

• ALPHA-Memory · In Development Measured Own sessions; method public only in outline. Internal measurement, not public. Checked 2026-10-04. Measured on one developer's recorded coding sessions against the memory system we used before. Not measured against any product in this table. The method is described on the home page only in outline; its harness and the recorded sessions are not public, so the measurement cannot be checked from outside.
• Claude Code Built-In Memory Unknown No measurement published in the memory docs. Vendor documentation, checked 2026-10-03.
• Supermemory Shipped Independent SWE-ContextBench on coding tasks; method public. uploaded to Supermemory’s managed cloud platform via client.add() Independent paper, checked 2026-10-04. SWE-ContextBench (arXiv 2602.08316) is not the vendor's own measurement; its method is public. It uploaded coding-agent sessions to Supermemory's hosted platform and retrieved through its MCP tools, not the Claude Code plugin. Supermemory's documentation cites it as "independent ones such as SWEContext" (supermemory.ai/docs/overview/what-is-supermemory).
• Hivemind Unknown Publishes LoCoMo cost, token and turn results. Vendor documentation, checked 2026-10-04.
• Mem0 Platform Unknown Independent SWE-ContextBench; ran Mem0 locally, not Platform. Each context is then committed to a local Mem0 Memory instance via two sequential add calls Independent paper, checked 2026-10-04. SWE-ContextBench (arXiv 2602.08316) is not the vendor's own measurement; its method is public. It stored coding-agent sessions in a local Mem0 instance with a local vector store, not in Mem0 Platform, and retrieved by querying that local vector store, so it does not settle the row for the hosted product. Mem0's research page (mem0.ai/research) lists conversational and tool-use benchmarks. Mem0 also published a two-experiment test inside Claude Code on its blog (mem0.ai/blog/how-mem0-cut-claude-code-s-memory-footprint-by-97, 23 July 2026) and says of it: "not a benchmark run".
• ContextStream Shipped Agentic coding-task benchmark; method section published. Each run started from a fresh clone and a fresh session. The prompts stayed identical. Hidden acceptance tests checked the work after the fact Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Unknown Publishes a code-exploration token benchmark only. Vendor documentation, checked 2026-10-03.
• Zep Cloud Unknown Publishes LoCoMo and LongMemEval results. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Unknown Benchmark suite on LoCoMo and BEAM datasets. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped SDE-bench coding bug-fix tasks; dataset public, method described. Deterministic pytest end to end; no LLM judge. Vendor documentation, checked 2026-10-04. The same post, line 126: "The dataset is public today".

Licence and Self-Hosting

Licence and self-hosting. Shipped: an open-source licence on the client or the engine, or a documented way to run the vendor's hosted store on infrastructure you control. No: neither. N/A: the memory the cell describes has no hosted store; the cell states its licence.

• ALPHA-Memory · In Development No Nothing released; open source not decided. This site: FAQ. Checked 2026-10-03.
• Claude Code Built-In Memory N/A Proprietary licence; auto memory stays on your machine. © Anthropic PBC. All rights reserved. Vendor documentation, checked 2026-10-04. The licence reserves all rights. Auto memory is files on the developer's machine, so it has no hosted store to self-host. Claude Code Projects (public beta on Pro and Max plans) have a project memory of their own, which the vendor calls "separate from the auto memory Claude Code keeps on your machine" (code.claude.com/docs/en/claude-projects).
• Supermemory Shipped Self-hosted binary, free within licence limit; open-source SDKs. Self-hosted is free within its lite license limit and useful for local development and privacy-sensitive workloads. The server binary is not open source Vendor documentation, checked 2026-10-04. The same page says: "The SDKs and other components in the public repository are open source; the downloadable self-hosted server binary is built from a separate, non-public codebase."
• Hivemind Shipped Apache-2.0 plugin and CLI; Hivemind Cloud default; BYOC. Apache License 2.0, © Activeloop, Inc. Vendor documentation, checked 2026-10-03.
• Mem0 Platform Shipped Apache 2.0 open-source engine and plugin; self-hosting documented. Apache License Version 2.0, January 2004 Vendor documentation, checked 2026-10-04. At the same commit: the plugin states the same licence in integrations/claude-code-plugin/.claude-plugin/plugin.json, line 10, and README.md, lines 138 to 150, documents the self-hosted server.
• ContextStream Shipped MIT client; hosted backend separate, not included. This repository contains the MIT-licensed Rust MCP server, client, editor setup, and release tooling. The hosted backend is separate and is not included here. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped Apache-2.0 engine, server, plugin; self-hosted server documented. The Apache-2.0 license applies to the open-source core, including the memory engine, Claude-Mem Server, CLI, SDKs, adapters, MCP tools, schemas, tests, examples, and public documentation. Vendor documentation, checked 2026-10-03.
• Zep Cloud Shipped BYOC for Enterprise; Community Edition deprecated; Graphiti open-source. For enterprise customers who need VPC residency and maximum control, we offer BYOC deployments where Zep runs in your own cloud infrastructure. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped AGPL-3.0 app; self-hosting offered. Prefer to self-host? The full app is free and open source (AGPL-3.0). Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped MIT repository licence; self-hosting documented. Hindsight is MIT licensed with no usage limits, no telemetry, and no restrictions. Vendor documentation, checked 2026-10-03.

SOC 2

SOC 2.

• ALPHA-Memory · In Development No No SOC 2 report. Our own statement; no report exists to inspect. Checked 2026-10-03.
• Claude Code Built-In Memory Shipped Anthropic SOC 2 Type 2 report (Trust Center). Learn more and access resources (SOC 2 Type 2 report, ISO 27001 certificate, etc.) at Anthropic Trust Center. Vendor documentation, checked 2026-10-04. Auto memory files stay on the developer's machine. The quote is from the Claude Code security page, which links the Anthropic Trust Center (trust.anthropic.com) for the report.
• Supermemory Shipped SOC 2 Type II; typically Scale and above. SOC 2 Type II | Certified | Independent audit of security controls. Available on production plans that advertise it (see pricing; typically Scale and above for formal enterprise packaging). Vendor documentation, checked 2026-10-03.
• Hivemind Shipped SOC 2 Type II; on the Enterprise plan. We are SOC 2 Type II certified, available on the Enterprise plan. Vendor documentation, checked 2026-10-04. The answer shows when the pricing page's question "What certifications do you have?" is opened. The page is headed "Hivemind Pricing"; its Enterprise card lists "SAML SSO, SOC 2, HIPAA".
• Mem0 Platform Shipped SOC-2 Type 1. SOC-2 Type 1 Vendor documentation, checked 2026-10-04.
• ContextStream Planned SOC 2 Type II audit underway. Type II audit underway. Scope and timeline available under NDA for prospective Enterprise customers. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Unknown No SOC 2 statement on its site. Vendor documentation, checked 2026-10-04. The home, pricing, privacy, security and terms pages of cmem.ai carry no SOC 2 statement (read on 2026-10-04).
• Zep Cloud Shipped SOC 2 Type II; reports for Enterprise subscribers. Zep’s managed service. SOC 2 Type II certified, with HIPAA Business Associate Agreements available for Enterprise customers. Vendor documentation, checked 2026-10-04. The same page says: "Compliance reports and supporting documentation are available to Enterprise subscribers." Zep's pricing page (https://www.getzep.com/pricing, read 2026-10-04) lists "SOC 2 Type II and HIPAA BAA" under its Enterprise plans.
• Basic Memory Teams No No own certification yet; platforms SOC 2 compliant. We don't yet hold our own SOC 2 certification. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped SOC2 Type 2 certified, per its product page. SOC2 Type 2 certified Vendor documentation, checked 2026-10-03.

Team Price

Team price published.

• ALPHA-Memory · In Development No No price; not for sale. This site: Contact. Checked 2026-10-03.
• Claude Code Built-In Memory N/A Part of Claude Code; no separate price. Vendor documentation, checked 2026-10-04. Claude Code is included in all paid Claude plans. The Team card on the pricing page lists "Claude Code and Claude Cowork" and prints a standard seat at "$20 Per seat / month if billed annually. $25 if billed monthly."
• Supermemory Shipped Pro $19 per month, 3 team seats. Pro $19 per month Get Pro $20 included per month For developers building with AI memory. Everything in Free, plus Auto top-up with configurable purchase limits 3 team seats, Vendor documentation, checked 2026-10-03. The quote is the Pro card of the pricing page, read top to bottom. The same page prints the Scale plan, "For teams and production workloads", at $399 per month.
• Hivemind Shipped $99 / seat / month, Team plan. $99 / seat / month Vendor documentation, checked 2026-10-04. The quote is the Team card. The same page prints the Basic plan at "$0 / month", pay as you go, with "Unlimited seats".
• Mem0 Platform Shipped Hobby Free; Starter $19/month; Pro $249/month; Enterprise Custom. Price Free $19/month $249/month Custom Vendor documentation, checked 2026-10-03.
• ContextStream Shipped $79 per user per month, Team, 3-seat minimum. Team is $79 per user per month with a 3-seat minimum: $237 per month for 600,000 pooled monthly credits. Vendor documentation, checked 2026-10-03.
• claude-mem / CMEM Cloud Shipped $333 per seat per month, Team Cloud. Team Cloud is $333 per seat per month with FDE onboarding. Vendor documentation, checked 2026-10-04. The quote is in the answer to the pricing page's question "What does cloud cost?", shown when the question is opened. The Team card on the same page prints "$333 / seat / mo".
• Zep Cloud Shipped Flex $125 per month; 50,000 credits included. Flex $125/ month billed monthly 50,000 credits included. Then $25 / 10,000 credits. Vendor documentation, checked 2026-10-03.
• Basic Memory Teams Shipped $15/seat/month Team; $13/seat at 10+, $12 at 20+. Team $15/seat/month Volume pricing: $13/seat at 10+, $12 at 20+. Vendor documentation, checked 2026-10-03.
• Hindsight Cloud Shipped Pay-as-you-go; no monthly fee, no per-seat pricing. Hindsight Cloud is pay-as-you-go. There is no fixed monthly fee and no per-seat pricing. Vendor documentation, checked 2026-10-03.

Products Not in the Table

Next to the memory built into Claude Code, which every team starts from, the table holds products that their vendor hosts. The tools below run on infrastructure you operate yourself, so they are not in it. The table is a selection, not a full list: other hosted memory products exist, and a product's absence here says nothing about it.

• Engram Local by default; Engram Cloud is self-hosted on infrastructure the user controls. Self-hosted cloud runtime: engram cloud serve on infrastructure you control Source, checked 2026-10-03.
• MemPalace Keeps everything on the user's machine by default; its team server is self-hosted. By default MemPalace keeps everything on your own machine. A central server is still your infrastructure Source, checked 2026-10-03.

Report a Correction

Is a cell wrong or out of date? Write to contact@alpha-memory.com and send the cell (product and row), a public link that shows the correct fact, and the date you saw it. A disputed cell reads Unknown until we have re-checked it.

What Changed

• 2026-10-04 First version: ALPHA-Memory and nine named products over 14 rows.

Product names are trademarks of their owners. ALPHA-Memory is not affiliated with, endorsed by or sponsored by any of them; the names are used only to identify the products compared.